
AI model
The AI model
First-contact AI, intent, sentiment, RAG and self-heal
Overview
Cusmato uses AI to support helpdesk staff. The goal is support, not replacement. The human remains responsible for the quality of the interaction.
First-contact AI
Automatic processing of incoming tickets
When a new ticket arrives: intent classification, sentiment analysis, RAG search, response generation. See the full pipeline below.
First-contact AI
When a new ticket arrives, it goes through the following steps:
1. Language detection
Detects the language of the customer message for the response.
2. Intent classification
refunddeliveryproduct_questioncomplaintother
3. Sentiment analysis
positive, neutral, negative — score -1 to 1.
4. Knowledge base (RAG)
Ollama embeddings, Qdrant, channelKey filtering.
5. Response generation
System prompt, onboarding data, RAG, tool calls, playbook.
6. Self-heal
Recovery on errors before escalation to human.
Jessica — AI assistant
Jessica supports staff with: ticket analysis, difference analysis, prompt corrections, learning rules. See Jessica for details.
Tools
| Tool | Description |
|---|---|
get_ticket_data | Retrieves ticket data from Redis |
get_all_outbox_tickets | Retrieves all outbox tickets |
get_tickets_with_feedback | Tickets with user feedback |
analyze_response_quality | Analyses response quality |
save_learning_rule | Saves learning rule |
get_learning_rules | Retrieves learning rules |
Quick-regenerate
When an agent gives feedback: feedback and rating are included, language is detected, improved response is generated.
What AI does
- Suggestions — Category, priority, intent, sentiment
- Draft response — First draft
- Knowledge base searches — Via RAG
- Language detection — Determine language
- Recovery on errors — Self-heal
What AI does not do
Limits
The agent always decides. No automatic sending without approval (unless configured). Closing a ticket requires human confirmation.
Trust and control
- Always verifiable — Every suggestion can be ignored or adjusted
- LLM approvals — Workflow for approval
- Feedback loop — Jessica and auto-improve process corrections
- Learning rules — Corrections in future prompts